Search papers, labs, and topics across Lattice.
The paper presents EnrgLLM, a large language model fine-tuned from Meta's Llama 3 using a question-answer dataset derived from the Society of Petroleum Engineers' content. This domain-specific fine-tuning aimed to improve the accuracy and relevance of LLM responses in the oil and gas domain. Subject matter expert evaluations demonstrated that EnrgLLM outperforms the base Llama 3 model in generating high-quality answers to petroleum engineering questions.
Forget generic LLMs: EnrgLLM proves fine-tuning Llama 3 on petroleum engineering data yields significantly better, expert-validated answers in the oil and gas domain.
This paper details the creation, testing, and deployment of EnrgLLM (referred to as EnergyLLM in earlier publications), a large language model (LLM) trained in the domain of petroleum engineering. The Society of Petroleum Engineers’ content was acquired and transformed into a question-answer training set and an opensource LLM, Llama 3 from Meta, was fine-tuned on this training set. The result is a language model that can generate higher quality answers about oil and gas questions than other LLMs. To validate this claim, we tested EnrgLLM vs. the base Llama model with subject matter experts. EnrgLLM scored higher than the base model according to their feedback. We then proceeded to prepare the deployment of EnrgLLM for a wider audience and connected it with the OnePetro corpus so it may cite sources in its answers. Each step of this process—the development of the dataset, training, testing, and deployment—are described herein.
